asm: Instructions file for Claude Code

CLAUDE.md

agent-skill-manager (asm) — a TypeScript CLI plus ink/React TUI that manages Agent Skills installed for Claude Code, Codex, Gemini, and OpenClaw.

Critical commands

npm install          # postinstall no-ops when CI or ASM_SKIP_POSTINSTALL is set
npm run build        # tsx scripts/build.ts -> dist/ (gitignored)
CI=true npm test     # unit suite — the command of record
npm run test:coverage  # v8 coverage for src/ — measurement only, no fail gate
npm run typecheck    # tsc --noEmit
npm run lint         # eslint over src/
npm run lint:site    # eslint over website-src/src
npm run test:e2e     # tests/e2e/ — NOT covered by `npm test`
npm start            # run the TUI from source

npm test is vitest run src/, and that argument is a substring filter on the test path, not a directory — website-src/src/__tests__/ matches it too, so npm run test:site is a subset of npm test, not a separate leg. Only tests/e2e/ is excluded. Measured timings and evidence: docs/AGENT_ENVIRONMENT.md.

npm run test:coverage is vitest run --coverage src/ with @vitest/coverage-v8 (5.x, matching current vitest). It reports line and branch percentages for src/ product files only (test files excluded). There is no coverage threshold and no CI fail gate — measurement before improvement (F-TEST-003 / #438). Vitest 5 remaps V8 coverage via AST, so the denominators below are not comparable to the vitest-4 #438 baseline; keep coverage.include.

Coverage baseline (2026-09-12, src/, v8, CI=true npm run test:coverage, 2800 tests):

• Lines: 82.69% (8740/10569)
• Branches: 73.74% (5393/7313)

src/commands/ measures 63.74% lines (#673): src/cli-*.test.ts dispatches runCLI() in-process via src/cli-test-harness.ts instead of spawning tsx children that v8 could not attribute (the layer previously read 10.53%). Still subprocess-bound by design: the runInlineTs readLine tests (stdin pipes need a real child), the cross-process borrow-lock test under asm get --path, and one bundle export no-overwrite case whose process.exit sits inside a catch that would swallow the in-process sentinel. toggle.ts (disable/enable) stays low because nothing exercises it — a real gap, not an attribution artifact.